No-Reference Image Quality Assessment Based on Human Visual System and Dual-Branch Multi-Level Residual Network

The objective of No-Reference Image Quality Assessment (NR-IQA) is to simulate the human visual system in evaluating image quality. The integration of the advantages of deep neural networks with the characteristics of the human visual system has been proven to be highly effective and accurate in numerous studies. This paper proposes a no-reference image quality assessment method based on the human visual system and a dual-branch multi-level residual network. Initially, the image is processed into HSV images that better align with human perception and Contrast-sensitivity weighted gradient images. These are then fed into two multi-level residual networks. The first network extracts multi-level content features from the HSV image, while the second network extracts multi-level structural and textural features from the Contrast-sensitivity weighted gradient image. Subsequently, the two multi-level features are fused to enhance the feature representation, with a weight module assigning different weights to the two features. Finally, these features are mapped to a quality score. Experimental results demonstrate that, on the five public datasets of LIVE, CSIQ, TID2013, LIVEC, and KonIQ-10k, the proposed method outperforms other advanced methods in terms of PLCC and SROCC metrics.